AI-powered drafting of formal insurance appeal letters for claim denials.
This repo re-packages an original one-shot Streamlit web app (which called the xAI cloud API) as an open, local-first Hermes skill. Any Hermes agent — on any model, fully on-device — can draft a complete, ready-to-send appeal letter conversationally. No cloud key required.
Give the agent the basic facts of a denial (your name, insurer, what was denied, why you need it), and it drafts a formal letter:
- Subject line
- Sender + insurer address blocks (line breaks preserved)
- Re: line with member / policy / denial details
- Medical-necessity argument
- 30-day response request
- Signature block
It never invents policy numbers, addresses, or clinical facts — placeholders
are flagged for you to fill in. See examples/sample-letter.md
for a full generated example (fabricated data).
hermes skills tap add https://github.com/steadfastgrowth/AdvocateAI
hermes skills list | grep advocateaiOr copy the repo and point Hermes at it:
hermes skills install /path/to/AdvocateAIThen in any Hermes session:
draft an appeal letter for my denied MS medication
The original called the xAI/Grok API and needed a browser form. A Hermes skill is:
- Local-first — runs on your on-box model (e.g. DeepSeek), no API key, no data leaving your machine.
- Model-agnostic — DeepSeek, Llama, Claude, Grok, anything Hermes supports.
- Conversational — collect the details in chat, get the letter back.
- Reusable craft — the drafting expertise is encoded as documented skill, not hidden in a UI.
| File | Purpose |
|---|---|
SKILL.md |
The skill: when to use, required fields, drafting prompt, output contract, pitfalls |
examples/sample-letter.md |
Full generated example letter (fabricated data) |
app.py |
Original Streamlit web form (kept for reference) |
requirements.txt |
Original deps (xAI SDK + streamlit) — not needed for the skill |
LICENSE |
MIT |
Open to PRs: sharper medical-necessity argumentation, new letter types (pre-auth, out-of-network, appeal of appeal), templates per insurer. Keep it local-first and model-agnostic.
MIT. See LICENSE.